The Aqua spacecraft is part of the NASA contri- bution to the Earth Observing System (EOS).
Aqua carries six state-of-the-art instruments to observe Earth oceans, atmosphere, land, ice and snow covers, and vegetation. The instruments provide high measurement accuracy, spatial detail and temporal frequency. A major poten- tial benefit from the Aqua and other EOS data is improved weather forecasting. Aqua, for exam- ple, carries a sophisticated array of instruments that will allow determination of atmospheric temperatures around the world to an accuracy of 1°C in 1-kilometre-thick layers throughout the troposphere, in conjunction with moisture profiles.
The capabilities of GOES systems have improved steadily, providing more rapid updating (time scales shorter than 30 minutes) and increasing spatial resolution. Rainfall estimation from GOES plat- forms uses infrared (IR) -based algorithms, based on the relationship between cloud-top growth and surface precipitation. The algorithms work best for convective rainfall and most poorly for shallow stratiform precipitation. The usefulness of polar- orbiting satellites for hydrologic forecasting, however, is severely limited by the frequency of overpasses of a given location. Combined systems of polar-orbiting satellites and GOES may provide important rainfall estimation capabilities for hydro- logical forecasting over the next few decades. These systems could be of particular utility for large river basins with poor raingauge networks and no radar coverage.
The potential for rainfall estimation from polar- orbiting platforms is well demonstrated by the
Tropical Rainfall Measuring Mission (TRMM) satel- lite (a NASA–Goddard Space Laboratory programme). The TRMM satellite contains a precip- itation radar, in addition to microwave and IR imagers. Various outputs of rain accumulations and potential flood and landslip impacts are available on the Website http://trmm.gsfc.nasa.gov. The examples in Figure 7.1 show three-hour rainfall accumulations and flood potential during a period of monsoon activity in northern Australia.
In a broader scope, international monitoring of rainfall (the Global Precipitation Measurement Missions (GPM)) has to be considered as a long- term programme lasting over several decades.
UNESCO has a collaboration agreement with WMO and the European Space Agency (ESA) on supporting these activities, in which other national and regional groups will become involved. As well as suitable satellite remote-sensing platforms, these facilities need effective ground truth from high-specification ground-data collection plat- forms. It is expected that solid-state scanning multi-frequency radar satellite instrumentation will be available to the meteorological community in general within the next decade. There are, however, concerns that this area of information collection will not maintain its advance. The number of working satellite sensors is declining and the development of sensor technology may also not be sufficient to meet the more detailed requirements needed for flood forecasting and warning. Although the continued development of EO for flood risk applications may be considered essential, national governments and international agencies are cutting back on investment in remote-sensing programmes.
Figure 7.1. (a) An example of a three-hourly output indicating rainfall accumulations in northern Australia for 3 February 2009 from the TRMM satellite. (b) TRMM satellite flood-risk information at the same
time as the rainfall information shown in (a)
CHAPTER 7. POTENTIAL APPLICATIONS FROM DEVELOPING TECHNOLOGIES 7-3
7.3 NUMERICAL WEATHER PREDICTION Major meteorological services are now using complex atmospheric–ocean models to produce NWP outputs to support their forecasting services.
The development of NWP has been closely allied to the evolution of remote-sensing and earth observa- tion. It has been widely successful in improving both the accuracy and lead time of weather fore- casts, and has also improved QPF. This section will examine the applicability of NWP to flooding and how it can be integrated into the flood forecasting and warning process.
7.3.1 Large-scale fluvial flooding
The weather systems that cause these events tend to be large scale and relatively predictable, often by several days in advance. Global wind, temperature and humidity profiles throughout the lowest 20 kilometres of the atmosphere, at 20 to 50 kilometres spatial resolution and 1-kilometre vertical resolution, are required every three to six hours to initialize NWP models. Observations of global sea temperature, ice and snow cover are also important inputs at a spatial resolution of approximately 5 kilometres. Increases in forecast accuracy are critically dependent on the observational data. Currently, existing satellites provide rather coarse vertical resolution of global temperature and humidity, but an improvement in resolution is likely in the next few years. Wind information at different levels is also available from satellite-based sensors.
7.3.2 Convective rainfall and small-scale
fluvial events
The weather systems that lead to these events are smaller scale and less predictable than the large- scale storms. Lead times will typically be in hours rather than days. The scale of forecasting therefore applies to the more rapidly responding catchments and also to pluvial and urban flooding.
NWP models for this type of flooding need to operate on a regional or local scale, and may require two steps of downscaling from the global models.
For north-west Europe the appropriate models use wind, temperature and humidity observations over the eastern North Atlantic and western Europe at 10- to 20-kilometres spatial resolution and 50– to 500-metres vertical resolution, throughout the lowest 20 kilometres of the atmosphere, every hour.
To capture the precursors to thunderstorm development, the same variables, together with cloud and precipitation, are also required over the United Kingdom every 15 to 60 minutes at 3- to 20-kilometres spatial resolution and 50- to 200-metres vertical resolution in the lowest
2 kilometres of the atmosphere. Sea and lake temperatures at 1-kilometre spatial resolution are required each day to optimize forecast accuracy. For most of these requirements, the dominant sources are currently satellite EO and ground-based radar, but in situ sources, such as commercial aircraft, also contribute significantly. Improvements in input data and their use can be expected to make substantial improvements in predictability of these events. Current satellite capabilities fall well short of requirements and effort is focused on upgrading radar capabilities to provide Doppler-based estimates of winds and humidity (through refraction measurements). For this purpose radars are required spaced less than 200 kilometres apart.
Most existing national radar networks have significant gaps and insufficient accuracy to meet this requirement.
To run NWP suites requires massive computer power, and along with the high level of data requirements, this is a high-cost, high-investment demand on governments. Only a few NMSs are capable of obtaining this level of support. Although Web-based products are available internationally, these can only, in reality, be used by a small NMS or flood warning service in an offline mode, to provide additional background information or early general alerts. There is scope, however, for individual RCMs and LAMs to be set up at national or basin scales, using data feeds and boundary conditions from an existing general circulation model (GCM).
7.4 GEOGRAPHICAL INFORMATION SYSTEMS
The use of GIS in a flood forecasting and warning system can provide a wide range of visualization products, containing far more information than basic mapping applications or text descriptions.
GIS use spatially related datasets and relational databases to provide successive layers or overlays of information, for example the identification of key infrastructure in relation to spatial flood risk and the movement of a flood through a channel or drainage system. GIS can provide data displays of conditions at given observation points for variables such as water level and rainfall. Commercial flood forecasting and warning software are generally available with a GIS interface, sometimes referred to as a graphical user interface (GUI). However, the GIS facility has to be “populated”, and gathering appropriate datasets is a significant undertaking in itself.
The most important dataset in relation to flood warning is the accurate representation of the land
surface by DTMs. These can be derived from digitizing contour maps, but will lack accuracy, especially in flood plains, which are areas of low relief. Photogrammetry from air photography has been largely superseded by digital airborne survey using either LIDAR (a laser-based scanning technology) or synthetic aperture radar (SAR), these being necessary to achieve the required levels of vertical accuracy, to 1 metre or less. LIDAR is considered more accurate than SAR. Some potential also exists for the quantification of roughness needed in flood-plain models from both airborne and satellite sensors.
High-resolution DTMs are a key requirement to provide accurate inundation forecast maps, as in the application by the Bangladesh FFWC for the city of Dhaka and its environs (Figure 7.2).
LIDAR is also being used to identify problematic flood-risk areas in constricted urban areas in the United Kingdom as part of a programme to develop a pluvial flood warning system (Making Space for Water, Defra, 2007). The accuracy of LIDAR is sufficient to identify small, low-lying areas where flood waters accumulate, and also to identify detailed flow paths. An example of this is shown in
Figure 7.3, which indicates applications on different scales relevant to the prediction of urban flooding.
7.5 INCORPORATING IMPROVEMENTS IN QUANTITATIVE PRECIPITATION FORECASTS
The main improvement provided by the use of QPFs is that they can provide objective numerical rainfall values over a given period, these replacing the subjective interpretations of commonly used terms for rainfall forecasts, such as “light”,
Figure 7.2. Inundation map for Dhaka based on a T+48-hour forecast
Source: FFWC of Bangladesh
Pseudo-quantitative rainfall forecast definitions (India and Bangladesh)
Qualitative description Rainfall amount (mm)
Light 4.57–9.64
Moderate 9.65–22.34
Moderately heavy 22.35–44.19
Heavy 44.20–88.90
Very heavy 89 +
CHAPTER 7. POTENTIAL APPLICATIONS FROM DEVELOPING TECHNOLOGIES 7-5
Figure 7. Localized flood depressions (a) and flow paths (b) identified by LIDAR for Carlisle and environs, United Kingdom
(a)
(b)
“moderate” and “scattered”. On the Indian subcontinent, most of the NMSs have a pseudo- quantitative version of descriptive terms, as shown in the table. The precision of numbers to two decimal places is unrealistic and it is not certain how these figures were originally derived. The classification is further limited in its usefulness in that there is no duration linked to the amount, so intensity is not recognized. The forecasts are usually applied to administrative areas, so the impact in the context of catchments is not directly relevant.
Obviously, such definitions as these will need to change between different climate types, and they also lack the necessary link with impacts. A 20-millimetre rainfall in one hour is insignifi- cant in Monsoon Asia, but a similar amount in a saturated small upland or highly urbanized catchment in a temperate zone could have signif- icant effects.
The main issues surrounding the use of QPFs in flood forecasting relate to how the information can be incorporated into models, the timeliness of their delivery and the confidence that can be placed upon the information.
Values of QPFs from meteorological models are normally provided in a pixel format, which is suit- able for inputs into grid-based models, but which would require areal and time integration if used in a lumped model. The spatial definition provided by the output depends on the type of meteorological model. It is considered that, for the pixel data to be usefully interpolated, a 3 × 3 array of pixels or a linear range of 5 pixels is required. Thus, at the scale of global climate models, definition is in tens of kilometres, which is not particularly useful except over the largest catchments. Post-event stud- ies in the United Kingdom have shown that detail at a level suitable for small catchment flood warn- ing requires a 4-kilometre model grid, as illustrated in Figure 7.4. A feature of NWP models (and hence QPF output) is that internal smoothing is required to maintain stability and accuracy, so the resolved scales are significantly coarser than the grid length.
It is considered (Golding, 2006) that a grid length of five units is required for accurate smoothing.
Thus, the 1.5-kilometre model should give good predictions for scales of 7.5 kilometres upwards, and a 4-kilometre model should give prediction on a 20-kilometre scale. It is clear from Figure 7.4 that the 12-kilometre resolution fails to identify both the high intensity and the localized nature of two major cells. These are, however, differentiated by the 1-kilometre model and are generally quite close to the recorded behaviour from radar.
An indication of the relative scales of spatial resolution and lead time for forecasting is shown in Figure 7.5. Errors in the location of a weather system in an NWP model depend on how it is being forced.
If the weather system is moving freely across the forecast domain, the error might be expected to increase at around 6 to 8 kilometres per hour, based on typical errors in wind speed at a height of a few kilometres above the surface. However, if the weather system is tied to the topography, the error should increase much more slowly. Analysis of
* Mesoscale convective system
Figure 7.5. Space scales of precipitation systems and their relationship with predictability
Hail shaft
Thunderstorm
front
Extra-tropical cyclone Space
scale
Lifetime Predictability Nowcast
10 mins 1 hr 12 hrs 3 days 30 mins 3 hrs 36 hrs 9 days 5 mins 30 mins 6hrs 36 hrs 1 000 km
100 km 10 km 1 km
MCS*
Figure 7.4. Comparison of rainfall accumulations from 1500 to 1700 hours universal coordinated time (UTC) for 3 August 2004. Note: the upper frames (a) and (b) are actuals from the radar
network, frame (c) shows the forecast at 0900 UTC from the 12-kilometre model and frame (d) shows forecast at 0900 UTC from the
1-kilometre model.
CHAPTER 7. POTENTIAL APPLICATIONS FROM DEVELOPING TECHNOLOGIES 7-7
development errors for individual thunderstorms suggests that individual storms should be predictable up to about three hours ahead.
Predictability refers to the identification of a forecastable feature. Nowcast refers to the ability to provide a detailed statement on behaviour, such as movement and QPF.
For extremely short lead times, there is not enough time to run a full NWP model, and the required resolution is too fine for current models in opera- tional mode. A cheaper alternative is to use linear extrapolation of recent radar observations. For individual thunderstorms, this approach can only provide accurate predictions for up to about half a storm lifetime, that is, 30 minutes, though for storms that are organized in a longer-lived line or group, there may be useful predictability for several hours ahead. Currently in the United Kingdom, STEPS is used. In addition to a “best esti- mate” QPF forecast, STEPS provides probabilistic information by perturbing the extrapolation vectors and by adding artificial variability at the small, unpredictable scales. The STEPS extrapola- tion forecast is merged with the 4-kilometre NWP model, taking advantage of the improved storm representation in this model. The positional accu- racy of a nowcasting system (forecasting between 0 and 12 hours ahead) depends predominantly on the extrapolation velocity. Given a 2-kilometre pixel size and a 15-minute interval between radar images, tracking is only likely to be accurate to 1 pixel in 15 minutes, or 8 kilometres per hour, though STEPS seeks to improve on this by combin- ing estimates from several neighbouring pixels.
Beyond half a storm lifetime, the development error dominates and only the overall movement of the group of storms can then be predicted by this method.
The error range inherent in QPFs has in some ways discouraged their use as direct inputs into hydro- logical and flood forecast models, and this is discussed more fully in 7.6. In practice, this has led to the use of QPFs in operational applications to provide an early stage of alert, or at most, to their use in forecast models run off-line to provide “what- if” scenarios. Their use to provide an early alert has provided the drive for better QPF information, which uses either probability-based or trigger-based forecasts, that is, forecast to exceed a predetermined threshold based on local conditions. The latter are more generally used, as they provide a single figure on which a decision linked to procedures can be made. Members of the engineering community, who are generally concerned with flood warning operations, have expressed reservations about using probability-based forecasts, as this requires judge- ments to be made in situations that are already difficult.
An extension of the use of QPF estimates is to convert the forecast into an estimate of runoff.
This projection does not necessarily require the intervention of a complex hydrological model, but is based on a water-balance updating approach. In the United Kingdom, the Met Office Surface Exchanges Scheme, incorporating a probability- distributed moisture model (MOSES-PDM) (Met Office and Centre for Ecology and Hydrology), is able to provide this within a DSF. In the United States, a similar method is used to provide flood alert information. The advantage of using a water balance approach is that it takes into account catch- ment conditions, whereas rainfall trigger values may be set at a low value to allow for the worst- case, saturated catchment conditions, which could create an undesirably high number of false alarms.
7.6 ASSESSMENT OF FORECASTING UNCERTAINTY AND HYDROLOGICAL ENSEMBLE PREDICTION
The issue of uncertainty in modelling has been dealt with extensively in Chapter 4. The purpose of revisiting the topic in this chapter is to consider some of the recent operational examples. All fore- casts contain uncertainty and one of the most successful ways of dealing with this has been the use of ensembles. The uncertainty associated with a hydrological forecast starts with the meteorology.
Given that all mesoscale atmospheric models attempt to model an essentially chaotic atmosphere this area has been the primary source of uncertainty for some years.
Whatever model is used, the presence of errors is inevitable and therefore must be built into a proba- bilistic forecast. Investigation of the subject indicates a very large disparity between the research community’s embrace of probabilistic forecasting (Fox and Collier, 2000; Journal of Hydrology, 2001 – Special Issue on Probabilistic and Ensemble Forecasting, Volume 249). The preference for a deterministic approach still prevails even in tech- nologically advanced hydrological forecasting agencies (see, for example, the Website of the California River Forecast Center). It must be assumed that overcoming this resistance will be a lengthy task and require considerable education, even where users can be assisted by providing deci- sion-support tools that make use of the probabilistic forecast (see 7.7). Whereas performance statistics, as discussed in Chapter 4, can be applied to quanti- ties of predicted flow against observed flow and the accuracy of timing (for example of flood peaks), event analysis also needs to look at the operational performance. The latter will include an examina- tion of the sequence and timing of weather forecasts
and warnings received, and whether these provided adequate response time, as well as the quantitative accuracy.
In 2006, the United Kingdom Met Office carried out a detailed examination of the quality of fore- casts and data supplied to the Environment Agency under the service agreement to support the latter agency’s flood forecasting and warning operations.
Heavy-rainfall warnings, QPFs, radar data feeds and storm, tide and surge forecasts were all subjected to a rigorous review to decide what items should be considered, how they could be assessed and how an automatic monitoring and evaluation process could be set up for data feeds (Met Office, 2006).
Developing the assessment programme highlighted the balance that has to be made between what it is desirable to measure and what is practical. This balance is perhaps best related to establishing what it is useful to know, as opposed to verifying all and every quantity just because a number is produced.
The main issues and outcomes are summarized below.
The Met Office review highlights the sheer volume of data handling required, both in a retrospective review of selected past data and in the planning of the subsequent operational phase. In the United Kingdom example, many products and data are electronically generated and distributed, and this in itself introduces issues of how data can be extracted for monitoring purposes. This becomes a signifi- cant issue when data extraction has to be in real time and has then to be available in formats that are appropriate to various manipulations. In deal- ing with retrospective data the review produced much less in the way of analytical results than had been anticipated. It was, however, able to test meth- ods and it produced useful guidance on what might or might not be appropriate in an operational context.
The availability and suitability of ground truth was shown to be a constraint in both the retrospective review, and for the operational phase of the project.
It was shown that forecast product development and monitoring requirements can be integrated to the benefit of both. Issues remain on what is appro- priate in terms of representativeness, sampling and the relationship between point and areal information.
The project used standard statistical methods for the analysis of results. As a general observation on the performance statistics, some of the measures produced highly variable results, which made it difficult to define what constitutes a desirable target. Thus, although it was proposed that bimonthly reporting of results should be applied to all products, results need to be reviewed in the
context of longer-term performance and most statistics should be presented as 12-month running means. Absolute values of performance statistics may have little meaning in isolation, but the behav- iour of a particular variable over time can illustrate changes in forecast quality. The project identified that, in some cases, assessment can be carried out using ranges or confidence limits and that this also provides a useful indication of forecast accuracy.
Establishing a broad-based assessment of a wide range of forecast products is not a trivial task. The recognition of this in the past may have influenced decisions on whether such a process was worth- while. Such an assessment is, however, of considerable benefit if it provides better means of reviewing performance in a more objective manner.
This is an advance on subjective judgements of individual forecasts, which lead to an impression of success or otherwise, rather than a factual measure.
7.7 OPERATIONAL USE OF FORECASTING UNCERTAINTY TO IMPROVE
DECISION-MAKING
Ensemble forecast techniques are beginning to be used for hydrological prediction by operational hydrological services throughout the world. These techniques are attractive because they allow the effects of a wide range of sources of uncertainty on hydrological forecasts to be accounted for.
Forecasting should not only offer an estimate of the most probable future state of a system, but also of the range of possible outcomes. Indeed, users are often more concerned with having a quantitative estimate of the probability that catastrophic effects may occur, than with knowing the most probable future state. Not only does ensemble prediction in hydrology offer a general approach to probabilistic prediction, but it also offers an approach to improve hydrological forecast accuracy.
International agencies such as the ECMWF have been investigating the use of MCS-based ensembles in recent years and a large-scale intercomparitive Hydrological Ensemble Prediction Experiment (HEPEX) has been studied since 2005. The main objective of HEPEX is to bring the international hydrological and meteorological communities together to demonstrate how to produce reliable
“engineering quality” hydrological ensemble fore- casts. The object is to produce forecasts that can be used with confidence to assist the water resources sector to make decisions that have important conse- quences for the economy and for public health and safety. Representatives of operational hydrological services and water resources agencies are expected